- Joined
- May 2026
- How they joined
- Directly
- Before Cognition
- QuoraSenior Staff Software Engineer, Product Team Lead
- Status
- Current
“Devin Fusion was the first major project I worked on after joining Cognition seven weeks ago. Over the weekend, I helped add Fable to the mix. … It was fascinating to see how much delegation styles vary across models, and how closely the best ones resemble good human management.”
- Devin Fusion (launched 2026-06-29) was his first major project at Cognition: in his words, "Devin Fusion was the first major project I worked on after joining Cognition seven weeks ago." Fusion pairs a frontier lead model with a cheaper sidekick model.Jul 13, 2026Devin Fusion was the first major project I worked on after joining Cognition seven weeks ago. Over the weekend, I helped add Fable to the mix. It made Devin smarter and, surprisingly, cheaper. It was fascinating to see how much delegation styles vary across models, and how closely the best ones resemble good human management. If this kind of work sounds interesting to you too, consider joining us at Cognition!
- Authored the Cognition blog post "Making Fable Cheaper Than Opus" (2026-07-13): swapping Opus 4.8 for Fable 5 as Devin's lead model lowered the bill despite Fable costing 2x per token, because Fable delegates earlier; Fable + sidekick cost $1.86/run vs $2.04 for Opus + sidekick on FrontierCode 1.1, scoring 60.7 vs 54.6.Jul 13, 2026Making Fable Cheaper Than Opus. By Joon Hee Lee. We replaced Opus 4.8 with Fable 5, and Devin's bill went down.
- Joined Cognition as Member of Technical Staff in May 2026 after nearly nine years at Quora; LinkedIn headline reads "Building Devin @ Cognition". He came from Quora/Poe directly, not via Cognition's Windsurf acquisition.May 2026
- At Quora (Mar 2024-May 2026) he was Senior Staff Software Engineer and product lead for Poe's Core Product team of 20+ engineers, designers and data scientists, working directly with the CEO (Adam D'Angelo) on Poe's strategy, owning all AI/ML features, LLM quality/evaluation, cost and speed, and third-party model provider relationships.Mar 2024
- Was the first engineer and tech lead manager on Poe's ML team (Staff ML Engineer, Aug 2022-Mar 2024) and tech lead manager for the Quora Feed team - so his routing/cost work at Cognition continues years of balancing LLM quality, latency and cost across many providers on Poe.Aug 2022
- Current role is Member of Technical Staff at Cognition.
- Spoke on "The State of Model Routing - NVIDIA, Cognition, OpenRouter" (AI Engineer, published 2026-08-06), positioning him as Cognition's public voice on multi-model routing.Aug 6, 2026The State of Model Routing - NVIDIA, Cognition, OpenRouter (AI Engineer)
- His Fable analysis ran 3,000 evaluation sessions across four configurations and found that in 81% of Fable-led runs the lead never made a code edit (vs 24% for Opus), and Opus made 4x more corrective edits at lead prices - evidence he works on agent delegation behaviour and eval methodology, not just routing infrastructure.Jul 13, 2026Making Fable Cheaper Than Opus. By Joon Hee Lee. We replaced Opus 4.8 with Fable 5, and Devin's bill went down.
- On Fusion he wrote that "Fable, for example, was much less of a micromanager than Opus or GPT 5.5. It also scoped the sidekick's work better and did a better job reviewing its code" and that "the best ones resemble good human management" - a management framing that echoes his years as a tech lead manager at Quora.Jul 13, 2026Devin Fusion was the first major project I worked on after joining Cognition seven weeks ago. Over the weekend, I helped add Fable to the mix. It made Devin smarter and, surprisingly, cheaper. It was fascinating to see how much delegation styles vary across models, and how closely the best ones resemble good human management. If this kind of work sounds interesting to you too, consider joining us at Cognition!
- Cognition's Devin Fusion launch post claims the multi-model harness keeps frontier and Fable 5-level performance at 60% lower cost on FrontierCode; Fusion uses lightweight classifiers for dynamic mid-session routing between the main agent and sidekick.Jun 29, 2026Devin Fusion launch: multi-model harness maintaining frontier performance at 60% lower cost on FrontierCode.
- Publicly recruits for Cognition: "If this kind of work sounds interesting to you too, consider joining us at Cognition!".Jul 13, 2026Devin Fusion was the first major project I worked on after joining Cognition seven weeks ago. Over the weekend, I helped add Fable to the mix. It made Devin smarter and, surprisingly, cheaper. It was fascinating to see how much delegation styles vary across models, and how closely the best ones resemble good human management. If this kind of work sounds interesting to you too, consider joining us at Cognition!
- Credited under Research in the acknowledgements of Cognition's FrontierCode 1.1 benchmark post, alongside Eric Lu, Ben Pan, Fermi Ma, Jacob Teo and Silas Alberti; FrontierCode is the benchmark his Fusion and Fable results are measured on.Jul 7, 2026
Product lead for Poe's Core Product team of 20+ product and AI engineers, designers, and data scientists. Work directly with the CEO to develop and execute Poe's product strategy and roadmap. Ownership areas include: • Core product features across all platforms (Web, iOS, Android) • All AI/ML-powered features on Poe • LLM quality & evaluation, costs, and speed • Interfacing with third-party model providers
• Tech lead manager for Quora Feed team • First engineer and tech lead manager on Poe ML team • Product team lead on Poe Core Product team
Tech lead for Feed team
Implemented a SiriKit extension for Google Wallet in Objective C to enable voice commands for sending and requesting payments. Created a general client for Wallet app extensions to communicate with the backend API and the local database. Designed and implemented new protocols and classes for Wallet SDK for use in other first-party apps.
Copenhagen Study Abroad Program (1 Semester)
Hacktech
Devin Fusion was the first major project I worked on after joining Cognition seven weeks ago. Over the weekend, I helped add Fable to the mix. It made Devin smarter and, surprisingly, cheaper. It was fascinating to see how much delegation styles vary across models, and how closely the best ones resemble good human management. If this kind of work sounds interesting to you too, consider joining us at Cognition!
Making Fable Cheaper Than Opus. By Joon Hee Lee. We replaced Opus 4.8 with Fable 5, and Devin's bill went down.
Devin Fusion launch: multi-model harness maintaining frontier performance at 60% lower cost on FrontierCode.
The State of Model Routing - NVIDIA, Cognition, OpenRouter (AI Engineer)
Really excited to launch this! You can now create and share custom AI apps on Poe through vibe coding - no coding knowledge required!
